Several agents: who owns the outcome
Parallel work, evidence and acceptance through a CI example
Slide contents
1. Several agents: who owns the outcome
Parallel work, evidence and acceptance through a CI example
2. Parallelism changes responsibility
Split work, pass evidence, accept the overall result
3. The lifecycle stays shared
Add branching to memory, recovery and outcome evaluation
4. Three reports are not an outcome
Atlas: plugin P upgraded, Node 22, commit c3
5. The authors compare team architectures
Kim and colleagues: Google Research, MIT, Google DeepMind
6. Messages do not define completion authority
Five architectures; different aggregation and stopping policies
7. The budget belongs to the system
Match resource ceilings; measure actual consumption separately
8. Define a successful attempt first
Task success, matched conditions, a known denominator
9. Decomposable research benefited from a team
Finance Agent: model-averaged success rate
10. Splitting sequential work can waste effort
PlanCraft: recipe, state and action depend on each other
11. Coding results need qualifications when transferred
SWE-bench Verified: a small subset and shared environment
12. The study yields no universal rule
Test decomposability rather than memorize a magic number
13. Failures emerge at role boundaries
Cemri and colleagues: 14 modes in three categories
14. The verifier also needs verification
Compiles is weaker than meets requirements
15. Delegation needs a checkable assignment
Objective, boundaries, return format and permitted actions
16. Send evidence with the conclusion
Keep artifacts, revisions and sources available to the coordinator
17. Shared memory does not resolve conflict
Isolated changes converge on one verifiable revision
18. Joining results needs explicit conditions
Replies alone are insufficient: check coverage and compatibility
19. The coordinator owns the decision
Workers propose; checks provide evidence; policy permits action
20. A branch retry must preserve intent
Repeating research and repeating an external effect differ
21. Evaluate the system and its trajectory
Task success, excess effects, cost and latency
22. Parallelize independent work, join dependent work
In CI, split investigation, then accept one revision
23. Next, choose the context that matters
Assignment contract, evidence, join and acceptance authority